{"id":121502,"date":"2025-09-10T15:44:00","date_gmt":"2025-09-10T07:44:00","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=121502"},"modified":"2025-09-11T00:45:24","modified_gmt":"2025-09-10T16:45:24","slug":"nvidia-blackwell-ultra%e5%9c%a8%e5%85%a8%e6%96%b0mlperf-inference%e5%9f%ba%e6%ba%96%e6%b8%ac%e8%a9%a6%e4%b8%ad%e6%a8%b9%e7%ab%8b%e6%96%b0%e6%a8%99%e7%ab%bf","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=121502","title":{"rendered":"NVIDIA Blackwell Ultra\u5728\u5168\u65b0MLPerf Inference\u57fa\u6e96\u6e2c\u8a66\u4e2d\u6a39\u7acb\u65b0\u6a19\u7aff"},"content":{"rendered":"\n<h3 class=\"wp-block-heading\">NVIDIA GB300 NVL72\u6a5f\u67b6\u898f\u6a21\u7cfb\u7d71\u5728\u5168\u65b0\u63a8\u7406\uff08reasoning\uff09\u63a8\u8ad6\u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0c\u5c55\u73fe\u6700\u9ad8\u8f38\u9001\u91cf\uff1b\u540c\u6642NVIDIA\u5e73\u53f0\u6301\u7e8c\u4fdd\u6301\u8457\u6240\u6709\u8cc7\u6599\u4e2d\u5fc3\u9805\u76ee\u7684\u300c\u6bcfGPU\u63a8\u8ad6\u300d\u7d00\u9304<\/h3>\n\n\n\n<p>\u63a8\u8ad6\u6548\u80fd\u81f3\u95dc\u91cd\u8981\uff0c\u56e0\u70ba\u5b83\u6703\u76f4\u63a5\u5f71\u97ff<a href=\"https:\/\/blogs.nvidia.com\/blog\/ai-inference-economics\/\" target=\"_blank\" rel=\"noreferrer noopener\">\u4eba\u5de5\u667a\u6167\uff08AI\uff09\u5de5\u5ee0\u7684\u7d93\u6fdf\u6548\u76ca<\/a>\u3002AI \u5de5\u5ee0\u57fa\u790e\u8a2d\u65bd\u7684\u8f38\u9001\u91cf\u8d8a\u9ad8\uff0c\u5c31\u80fd\u4ee5\u9ad8\u901f\u7522\u751f\u66f4\u591a<a href=\"https:\/\/blogs.nvidia.com\/blog\/ai-tokens-explained\/\" target=\"_blank\" rel=\"noreferrer noopener\">\u8a5e\u5143\uff08token\uff09<\/a>\uff0c\u9032\u800c\u63d0\u9ad8\u71df\u6536\u3001\u964d\u4f4e\u7e3d\u64c1\u6709\u6210\u672c\uff08TCO\uff09\uff0c\u4e26\u63d0\u5347\u6574\u9ad4\u7cfb\u7d71\u7684\u751f\u7522\u529b\u3002<\/p>\n\n\n\n<p>\u63a1\u7528 <a href=\"https:\/\/developer.nvidia.com\/blog\/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era\/\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA Blackwell Ultra \u67b6\u69cb<\/a>\u7684 NVIDIA GB300 NVL72 \u6a5f\u67b6\u898f\u6a21\u7cfb\u7d71\uff0c\u5728 NVIDIA GTC \u5927\u6703\u9996\u5ea6\u4eae\u76f8\u81f3\u4eca\u4e0d\u5230\u534a\u5e74\u5f8c\uff0c\u5df2\u5728 MLPerf Inference v5.1 \u7684\u5168\u65b0\u63a8\u7406\uff08reasoning\uff09\u63a8\u8ad6\u57fa\u6e96\u4e0a\u5275\u4e0b\u7d00\u9304\uff0c\u8207\u63a1\u7528 NVIDIA Blackwell \u67b6\u69cb\u7684 GB200 NVL72 \u7cfb\u7d71\u76f8\u6bd4\uff0cDeepSeek-R1 \u63a8\u8ad6\u7684\u8f38\u9001\u91cf\u6700\u9ad8\u53ef\u9054 1.4 \u500d\u3002<\/p>\n\n\n\n<p>Blackwell Ultra \u5efa\u7acb\u5728 Blackwell \u67b6\u69cb\u7684\u6210\u529f\u4e4b\u4e0a\u3002Blackwell Ultra \u67b6\u69cb\u63d0\u4f9b\u8207 Blackwell \u76f8\u6bd41.5 \u500d\u7684 NVFP4 AI \u904b\u7b97\u8207 2 \u500d\u7684 attention-layer \u52a0\u901f\uff0c\u4e14\u6bcf\u9846 GPU \u53ef\u914d\u5099\u6700\u9ad8 288GB \u7684 HBM3e \u8a18\u61b6\u9ad4\u3002<\/p>\n\n\n\n<p>NVIDIA \u5e73\u53f0\u4e5f\u5728 MLPerf Inference v5.1 \u5957\u4ef6\u65b0\u589e\u7684\u6240\u6709\u8cc7\u6599\u4e2d\u5fc3\u57fa\u6e96\u4e0a\u5275\u4e0b\u6548\u80fd\u7d00\u9304\uff0c\u5305\u542b DeepSeek-R1\u3001Llama 3.1 405B Interactive\u3001Llama 3.1 8B \u8207 Whisper\uff0c\u540c\u6642\u6301\u7e8c\u7dad\u6301\u6bcf\u9846 GPU \u5728\u6bcf\u9805 MLPerf \u8cc7\u6599\u4e2d\u5fc3\u57fa\u6e96\u4e2d\u7684\u7d00\u9304\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1200\" height=\"675\" src=\"https:\/\/www.ioiotimes.com\/wordpress\/wp-content\/uploads\/2025\/09\/20250910-nvidia02.png\" alt=\"\" class=\"wp-image-121670\" title=\"\" srcset=\"https:\/\/www.ioiotimes.com\/wordpress\/wp-content\/uploads\/2025\/09\/20250910-nvidia02.png 1200w, https:\/\/www.ioiotimes.com\/wordpress\/wp-content\/uploads\/2025\/09\/20250910-nvidia02-300x169.png 300w, https:\/\/www.ioiotimes.com\/wordpress\/wp-content\/uploads\/2025\/09\/20250910-nvidia02-1024x576.png 1024w, https:\/\/www.ioiotimes.com\/wordpress\/wp-content\/uploads\/2025\/09\/20250910-nvidia02-768x432.png 768w\" sizes=\"(max-width: 1200px) 100vw, 1200px\" \/><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5168\u7aef\u6574\u5408<\/strong><strong><\/strong><\/h3>\n\n\n\n<p>\u5168\u7aef\u5171\u540c\u8a2d\u8a08\u5728\u4ea4\u51fa\u9019\u4e9b\u6700\u65b0\u57fa\u6e96\u6210\u7e3e\u6642\u626e\u6f14\u91cd\u8981\u89d2\u8272\u3002Blackwell \u8207 Blackwell Ultra \u5c0e\u5165\u5c0d NVFP4 \u8cc7\u6599\u683c\u5f0f\u7684\u786c\u9ad4\u52a0\u901f\u3002NVFP4 \u662f NVIDIA \u8a2d\u8a08\u7684 4 \u4f4d\u5143\u6d6e\u9ede\u683c\u5f0f\uff0c\u76f8\u8f03\u65bc\u5176\u4ed6 FP4 \u683c\u5f0f\u63d0\u4f9b\u66f4\u4f73\u6e96\u78ba\u5ea6\uff0c\u4e5f\u53ef\u9054\u5230\u8207\u8f03\u9ad8\u7cbe\u5ea6\u683c\u5f0f\u76f8\u7576\u7684\u6e96\u78ba\u5ea6\u3002<\/p>\n\n\n\n<p>NVIDIA TensorRT Model Optimizer \u8edf\u9ad4\u5c07 DeepSeek-R1\u3001Llama 3.1 405B\u3001Llama 2 70B \u8207 Llama 3.1 8B \u91cf\u5316\u70ba NVFP4\u3002\u642d\u914d\u958b\u6e90\u7684 NVIDIA TensorRT-LLM \u7a0b\u5f0f\u5eab\uff0c\u9019\u9805\u6700\u4f73\u5316\u4f7f Blackwell \u8207 Blackwell Ultra \u5728\u63d0\u4ea4\u6210\u7e3e\u6642\u80fd\u5728\u6eff\u8db3\u56b4\u683c\u6e96\u78ba\u5ea6\u8981\u6c42\u7684\u540c\u6642\u63d0\u4f9b\u66f4\u9ad8\u6548\u80fd\u3002<\/p>\n\n\n\n<p>\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u63a8\u8ad6\u7531\u5169\u7a2e\u57f7\u884c\u7279\u6027\u622a\u7136\u4e0d\u540c\u7684\u5de5\u4f5c\u8ca0\u8f09\u7d44\u6210\u3002\u9019\u5305\u62ec\u7528\u65bc\u8655\u7406\u4f7f\u7528\u8005\u8f38\u5165\u4ee5\u7522\u751f\u7b2c\u4e00\u500b\u8f38\u51fa\u8a5e\u5143\u7684\u60c5\u5883\uff08context\uff09\uff0c\u4ee5\u53ca\u7528\u65bc\u7522\u751f\u5176\u5f8c\u7684\u6240\u6709\u8f38\u51fa\u8a5e\u5143\u7684\u751f\u6210\uff08generation\uff09\u3002<\/p>\n\n\n\n<p>\u4e00\u7a2e\u7a31\u70ba\u300c\u5206\u96e2\u5f0f\u670d\u52d9\uff08disaggregated serving\uff09\u300d\u7684\u6280\u8853\uff0c\u6703\u5c07\u60c5\u5883\u8207\u751f\u6210\u4efb\u52d9\u62c6\u5206\uff0c\u4f7f\u6bcf\u4e00\u90e8\u5206\u90fd\u80fd\u7368\u7acb\u6700\u4f73\u5316\u4ee5\u9054\u5230\u6700\u4f73\u6574\u9ad4\u8f38\u9001\u91cf\u3002\u9019\u9805\u6280\u8853\u662f Llama 3.1 405B Interactive \u57fa\u6e96\u5275\u4e0b\u7d00\u9304\u7684\u95dc\u9375\u4e4b\u4e00\u3002\u76f8\u8f03\u65bc\u5728 NVIDIA DGX B200 \u4f3a\u670d\u5668\u4e0a\u4ee5\u50b3\u7d71\u670d\u52d9\u67b6\u69cb\u57f7\u884c\u8a72\u57fa\u6e96\u6642\u6bcf\u9846 Blackwell GPU \u7684\u8868\u73fe\uff0cGB200 NVL72 \u7cfb\u7d71\u7684\u6bcf\u9846 GPU \u6548\u80fd\u63d0\u5347\u5c07\u8fd1 50%\u3002<\/p>\n\n\n\n<p>NVIDIA \u4e5f\u5728\u672c\u8f2a\u57fa\u6e96\u6e2c\u8a66\u9996\u5ea6\u4f7f\u7528 <a href=\"https:\/\/www.nvidia.com\/zh-tw\/ai\/dynamo\/\" target=\"_blank\" rel=\"noreferrer noopener\">NVIDIA Dynamo \u63a8\u8ad6\u6846\u67b6<\/a>\u63d0\u4ea4\u6210\u7e3e\u3002<\/p>\n\n\n\n<p>NVIDIA \u7684\u5408\u4f5c\u5925\u4f34\uff0c\u5305\u62ec\u96f2\u7aef\u670d\u52d9\u4f9b\u61c9\u5546\u8207\u4f3a\u670d\u5668\u88fd\u9020\u5546\uff0c\u63d0\u4ea4\u4e86\u63a1\u7528 NVIDIA Blackwell \u8207\uff08\u6216\uff09Hopper \u5e73\u53f0\u6240\u5c55\u73fe\u7684\u4eae\u773c\u6210\u7e3e\u3002\u9019\u4e9b\u5925\u4f34\u5305\u542b Azure\u3001Broadcom\u3001Cisco\u3001CoreWeave\u3001Dell Technologies\u3001\u6280\u92fc\u79d1\u6280\u3001HPE\u3001Lambda\u3001Lenovo\u3001Nebius\u3001Oracle\u3001\u96f2\u9054\u79d1\u6280\u3001Supermicro \u53ca University of Florida\u3002<\/p>\n\n\n\n<p>\u5728 NVIDIA AI \u5e73\u53f0\u4e0a\u9818\u5148\u5e02\u5834\u7684\u63a8\u8ad6\u6548\u80fd\uff0c\u73fe\u5df2\u7531\u4e3b\u8981\u96f2\u7aef\u670d\u52d9\u4f9b\u61c9\u5546\u8207\u4f3a\u670d\u5668\u88fd\u9020\u5546\u63d0\u4f9b\u3002\u9019\u70ba\u90e8\u7f72\u9032\u968e AI \u61c9\u7528\u7684\u7d44\u7e54\u5e36\u4f86\u66f4\u4f4e\u7684\u7e3d\u64c1\u6709\u6210\u672c\u8207\u66f4\u9ad8\u7684\u6295\u8cc7\u5831\u916c\u3002<\/p>\n\n\n\n<p>\u82e5\u60f3\u9032\u4e00\u6b65\u4e86\u89e3\u9019\u4e9b\u5168\u7aef\u6280\u8853\uff0c\u8acb\u95b1\u8b80 NVIDIA \u6280\u8853\u90e8\u843d\u683c<a href=\"https:\/\/developer.nvidia.com\/blog\/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut\/\" target=\"_blank\" rel=\"noreferrer noopener\">\u95dc\u65bc MLPerf Inference v5.1 \u7684\u6587\u7ae0<\/a>\u3002\u4e5f\u6b61\u8fce\u9020\u8a2a<a href=\"https:\/\/docs.nvidia.com\/dgx-cloud\/benchmarking\/performance-explorer-user-guide\/latest\/accessing.html\/\" target=\"_blank\" rel=\"noreferrer noopener\"> NVIDIA DGX Cloud Performance Explorer<\/a>\uff0c\u4ee5\u6df1\u5165\u8a8d\u8b58 NVIDIA \u7684\u6548\u80fd\u8207\u6a21\u578b\u7e3d\u64c1\u6709\u6210\u672c\uff0c\u4e26\u7522\u751f\u5ba2\u88fd\u5316\u5831\u544a\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u3000\u3000\u3000\u3000<\/p>\n\n\n\n<h4 class=\"wp-block-heading has-text-align-right has-very-light-gray-to-cyan-bluish-gray-gradient-background has-background\">\ud83d\udfe6<a href=\"https:\/\/reurl.cc\/x3o32z\" target=\"_blank\" data-type=\"link\" data-id=\"https:\/\/reurl.cc\/x3o32z\" rel=\"noreferrer noopener\">ASUS T500\u96fb\u7af6\u684c\u6a5f \u770b\u6587\u7ae0\u5f97\u597d\u79ae <\/a><br>\ud83d\udfe6<a href=\"https:\/\/reurl.cc\/K9Ol2n\" target=\"_blank\" rel=\"noreferrer noopener\">be quiet!  \u8d85\u70ab\u6d77\u666f\u4e3b\u6a5f \u770b\u6587\u7ae0\u5f97\u597d\u79ae <\/a><br>\ud83d\udfe6\u73fe\u5728\u5c31\u52a0\u5165&nbsp;ioioTIMES \u81c9\u66f8\u7c89\u7d72\u5718&nbsp;\u66f4\u591a\u4e92\u52d5\u3001\u66f4\u591a\u597d\u5eb7\u650f\u62b5\u52a0!!<br>\ud83d\udfe6<strong>\u6211\u5011\u6709<a href=\"https:\/\/today.line.me\/tw\/v2\/publisher\/103117\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">LINE TODAY<\/a>\u983b\u9053\u4e86\uff0c\u5feb\u4f86\u8ffd\u8e2a\u6211\u5011\u5427!!&#8211;\u6700\u65b0\u79d1\u6280\u65b0\u805e \u76e1\u5728\u4f60<\/strong>\u624b<br><\/h4>\n","protected":false},"excerpt":{"rendered":"<p>NVIDIA GB300 NVL72<\/p>\n","protected":false},"author":3,"featured_media":121670,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[13],"tags":[13024,10805,71,15029],"class_list":["post-121502","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-blackwell-ultra","tag-mlperf-inference","tag-nvidia","tag-gpu"],"_links":{"self":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/121502"}],"collection":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=121502"}],"version-history":[{"count":2,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/121502\/revisions"}],"predecessor-version":[{"id":121671,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/121502\/revisions\/121671"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/media\/121670"}],"wp:attachment":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=121502"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=121502"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=121502"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}